A new open-source speech-to-text application for macOS reached the front page of Hacker News on April 6, 2026, offering developers a privacy-focused alternative to cloud-based voice input tools. Ghost Pepper, built by developer MattHart88, performs all speech recognition and text processing locally on the user's machine, ensuring no voice data leaves the device.
Privacy-First Design with Multiple Model Options
Ghost Pepper operates as a menu bar application that users activate by holding the Control key to record and releasing to transcribe. The tool integrates WhisperKit for speech recognition, offering four model options ranging from the lightweight Whisper tiny.en (75 MB) to the multilingual Parakeet v3 (1.4 GB) supporting 25 languages via FluidAudio.
The application uses Qwen LLMs for text cleanup, removing filler words and correcting self-corrections. Users can choose between three Qwen models: the default 0.8B version (535 MB) completes processing in 1-2 seconds, while the 4B model (2.8 GB) offers highest quality at 5-7 seconds processing time.
Strong Community Adoption and Active Development
The GitHub repository accumulated 526 stars and 25 forks shortly after launch, with 12 releases demonstrating active development. The creator posted on Hacker News that he built the tool to explore "how far I could get with a voice-to-text app that used 100% local models" and has been using it extensively for coding and emails.
MattHart88 is also experimenting with Ghost Pepper as a voice interface for AI agents, suggesting potential applications beyond simple transcription. The project is released under the MIT license, inviting community contributions and extensions.
Technical Implementation Prioritizes Security
Ghost Pepper's privacy architecture ensures transcriptions are never written to files and the application runs without a dock icon for minimal system footprint. The hold-to-talk interface allows users to paste transcribed text directly into any text field across macOS applications.
The tool represents a growing trend of on-device AI solutions for privacy-sensitive workflows, particularly among developers concerned about sending voice data to cloud services. With multiple model options balancing speed and accuracy, users can optimize the tool for their specific hardware and use cases.
Key Takeaways
- Ghost Pepper performs all speech recognition locally on macOS using WhisperKit and Qwen LLMs, with no data leaving the device
- The application offers four speech recognition models ranging from 75 MB to 1.4 GB, supporting both English-only and multilingual transcription
- The project reached 216 points on Hacker News and accumulated 526 GitHub stars shortly after launch on April 6, 2026
- Users activate transcription by holding Control to record and releasing to paste text directly into any application
- The open-source MIT-licensed tool is being used by its creator for coding, emails, and experimental voice interfaces for AI agents